Estimating above ground tree biomass for the Uncompahgre Plateau in Western Colorado using NAIP imagery and a series of textural and probabilistic metrics

Background/Question/Methods

           Detailed, accurate, efficient, and inexpensive methods of estimating basal area, trees, and aboveground biomass per acre (BAA, TPA, and AGB respectively) across broad scales are needed to effectively manage forests.  We developed such a methodology that uses readily available National Agriculture Imagery Program (NAIP) imagery, Forest Inventory Analysis (FIA) field plot data, and a novel two-stage classification and estimation approach. The first stage of our technique estimates the probability that each meter² pixel within the 580,000-acre Uncompahgre Plateau (UP) represents each of 15 cover types using visually interpreted NAIP stratified samples, a principle component analysis (PCA) of the NAIP imagery, multiple Gray Level Co-occurrence Matrix (GLCM) values of the first and third principle components, and polytomous logistic regression (PLR). The second stage of our procedure uses multivariate regression (MR) and the spatial locations of the FIA plot data to relate BAA, TPA, and AGB measurements to metrics derived from our first stage outputs.   

Results/Conclusions

            The top ranking PLR model accurately predicts cover type probabilities (Max re-scaled R² = 0.8220) and is statistically significant (p-value < 0.0001). Predictor variables that contribute to the model include three principle components (interpreted as brightness, neighboring pixel variation, and a vegetative index) and multiple GLCM values for a 3 by 3 moving window using the first and third components. Combining these probabilities in the second stage of analysis, we created a series of predictive surfaces that quantify multiple aspects of tree canopy edge and cover.

            The top ranking MR model is statistically significant (p-values < 0.0001) and accurately estimates BAA, TPA, and AGB for six dominant tree species sampled within the study region (R², 0.56 – 0.70). Covariates that contribute to this model include multiple metrics of species canopy area, shape, edge, and texture summarized to match the spatial foot print of the FIA plot data. Using this model we created a series of raster surfaces depicting mean BAA, TPA, and AGB across the UP. These surfaces provide UP land managers and others with inexpensive and detailed information that can be easily manipulated within a geographic information system to answer a wide range of questions.